In the analysis of X-ray images for security screening, challenges such as object overlap and occlusion, interference from similar materials and shapes, resolution limitations, and restricted viewing angles have consistently posed significant difficulties for the accurate detection of hazardous items. To address the issues of occlusion and small target detection in X-ray images, this paper proposes a hazardous object detection method based on YOLOv10 and the Efficient Multi-Head Multi Attention (EMMA) mechanism. Initially, EMMA is integrated into the C2f component of YOLOv10, employing adaptively weighted channel features to enhance the capability of capturing multi-scale features, thereby improving detection performance for targets of varying sizes, particularly for occluded and small objects. Subsequently, in the backbone of YOLOv10, the multi-head attention ESHA within the self-attention module (PSA) is replaced with the proposed EMMA, achieving an optimal balance between speed and performance in hazardous object detection on X-ray images. A dynamic focusing bounding box loss function, Focaler-IoU, is then introduced to dynamically adjust bounding boxes and enhance the model’s ability to detect targets across different scales. Finally, ablation experiments and comparisons with other detection algorithms are conducted using the publicly available OPIXray dataset. The results demonstrate that the improved model achieves an average precision (mAP@0.5) of 83.2%, representing a 3.0% improvement over the official YOLOv10n. Additionally, Grad-CAM is utilized to generate heatmaps that reveal the areas of the image the model focuses on during inference. The experimental outcomes indicate that this method effectively enhances the detection of occluded and small hazardous objects in X-ray images, meeting the accuracy and real-time detection requirements in complex security screening scenarios.